← Back to perspectives
/Intellicon FDE Team

What Should You Do When AI Speeds Up Development but Security Reviews Can't Keep Up?

Surveyed teams checked transitive dependencies in only about 53% of cases. Before expanding the use of AI development tools, take stock of changes awaiting review and find out why fixes have been deferred.

AIEgentWrX人機協作導入
What Should You Do When AI Speeds Up Development but Security Reviews Can't Keep Up?

Development teams are using AI to deliver new features while security staff are still investigating the previous batch of alerts. Project leads are caught in the middle: business units are waiting to go live, but reviewers cannot assess the risks before the deadline. Before deciding whether to expand AI use, we recommend checking how many code changes still need review and where outstanding fixes stand. Faster coding alone is not a reason to schedule more development work.

On November 13, 2026, iThome reported on a Linux Foundation survey of 308 security professionals. Among respondents, 58% said software delivery had become too fast for security reviews to keep up, and 60% said their code required more external dependencies after they started using AI development tools. The study did not report separate findings for Taiwan or manufacturing, so these percentages cannot be used to estimate conditions at your own company. However, managers can start by asking: how much of the new code still hasn't been reviewed?

Find Out Where Work Is Backing Up Before Increasing Development Volume

Development and security leaders can start with one project that uses AI development tools and jointly identify code changes awaiting review, alerts that have not been investigated, and issues whose fixes have been deferred. Each item should include a named owner, the date it was raised, a deadline, and the reason it is currently blocked. Listing a department is not enough if no one is responsible for determining the next step.

For example, when a procurement price-comparison system is about to add a feature for importing supplier quotes, the project lead can ask the team to list three categories separately: completed code changes that have not been reviewed, alerts from security tools that have not been investigated, and confirmed vulnerabilities that have not been fixed. These categories require different kinds of support and should be tracked separately, rather than grouped together as unfinished development work.

During weekly progress reviews, check more than the number of open items. Confirm how long the oldest items have been waiting, whether critical issues have missed their deadlines, and whether anyone has verified completed fixes. If the backlog keeps growing, we recommend reducing the number of code changes submitted for review at the same time, then adjusting staffing or supplying missing information based on what is causing the delays. Simply extending deadlines may remove overdue items from the list, but it does not resolve the underlying issues.

Development and security leaders should agree in advance on the conditions for expanding AI use. For example, they might require that all critical alerts have an action plan, fixes are being completed as agreed, and review wait times are no longer increasing before allowing more projects to use AI.

Before Adding a Package, Ask Whether Existing Tools Can Do the Job

The same survey found that teams checked transitive dependencies in only about 53% of cases. These are the additional packages that must be included for a newly added package to work. The researchers recommend reducing unnecessary dependencies and obtaining third-party packages from vetted repositories that comply with the organization's security policies. Managers can therefore start by asking: does this feature really require another external package?

We recommend that development and security teams jointly create a checklist for adding packages. Developers should explain what the package is for, whether an existing package could serve the same purpose, where it comes from, and who will maintain it. They should also provide lists of direct and transitive dependencies. These lists and the vulnerability scan results should come from existing development and security tools, rather than from someone guessing based on package names.

If your company uses EgentWrX, you can define the rules for preparing the checklist as a Skill, so the AI Agent follows the same rules each time. Using the information staff provide, the AI Agent identifies missing content and items that still need confirmation, then passes them to the responsible person for a decision. Before the Skill can be made available company-wide, it must be submitted to the review queue in the admin console and approved.

For example, if the information provided does not include a list of transitive dependencies, it should be marked as missing rather than treated as evidence that no other packages are involved. The requester should first complete the details on the package's purpose and source, then submit them to security staff for an assessment of whether the package is acceptable. This avoids sending incomplete requests straight into review.

Let AI Organize Review Materials, While People Remain Responsible for Approval and Fixes

We recommend separating information preparation from professional judgment. An AI Agent can help compile change descriptions, alert lists, and handling records provided by staff. Security personnel should then confirm whether an issue exists, what needs to be addressed first, and whether the fixes meet requirements. Having the information organized does not mean the code is ready to go live.

With EgentWrX tasks and workflows, you can break the preparation work into repeatable tasks and connect them in the order they need to run. When configuring a workflow, add a human approval step between consecutive tasks. Once the first task has prepared the materials for review, a designated person checks them against the source information, identifies omissions and unresolved questions, and confirms that they are ready before the next task proceeds.

This approval step determines whether information in the workflow can be passed to the next task. Actual vulnerability scanning, fix verification, and code releases still need to be handled by the responsible people using existing tools. Project leads must therefore also confirm the release approval process in the development system, so the team does not mistake permission to continue a preparation task for permission to release the code.

For deferred fixes, we recommend recording more than “concerns about stability” or “not enough time.” Specify which functions need testing, who will arrange the tests, and when the remediation date will be reviewed. With this information, managers can use the next meeting to assess what information and support are still needed, decide whether to adjust development volume, allocate testing resources, or add security reviewers, and then determine whether to expand AI use.

FAQ

Should we stop using AI development tools if security reviews can't keep up?

We recommend pausing further expansion and reviewing the status of existing projects, rather than immediately disabling every tool. Development and security leaders should first identify what is causing the backlog. If critical alerts remain uninvestigated and fixes are repeatedly overdue, reduce the number of code changes submitted for review at the same time and assign people to address the outstanding work.

When adding a package, is it enough to check only the packages developers use directly?

No. Both direct and transitive dependencies should be checked, because a new package may require other packages to function. Ask developers to provide dependency lists and vulnerability scan results generated by existing tools, along with the package's source and the person responsible for maintaining it, then submit that information to security staff for assessment.

Can managers approve a release based solely on a summary prepared by an AI Agent?

We do not recommend it. Summaries prepared by an AI Agent still need to be checked by security personnel. Reviewers should verify the original alerts, the code changes, and the results of fix verification before deciding whether to approve the release under the company's release procedures. Organized information should not be mistaken for a completed security review.

References

30 minutes to map out which work to hand to AI first

Want every employee to have their own AI teammate?

A consultant will be in touch shortly.

ISO/IEC 27001:2022 Certified ISMS